To Answer or Not to Answer? Improving Machine Reading Comprehension Model with Span-based Contrastive Learning